Evaluating Machine Learning Models for Predict Cost Overruns in Petrochemical Projects

Predictive analytics are important in managing cost overrun classification in petrochemical projects. Accurate prediction models are essential for effective project management, allowing informed decision-making and risk mitigation. This paper comprehensively evaluates five machine learning classification models (RF, DT, NB, SVM, and ANN) to predict cost overruns in petrochemical projects. The models were assessed using Test/Train split and K-fold cross-validation methods, focusing on key metrics such as Precision (PR), Recall (RC), F1-Score(F1), and Accuracy (ACC). Naive Bayes demonstrated superior performance in the Test/Train split evaluation, achieving up to 76% accuracy, particularly in the 25/75 splits. It consistently balanced PR and RC, making it the most reliable model for unseen data. In contrast, the Decision Tree model excelled in the K-fold cross-validation, reaching 96% accuracy and a 0.95 F1 at K-fold = 10, proving its effectiveness across diverse datasets. Overall, Naive Bayes emerged as the best Test/Train split evaluation model, while Decision Tree was the top performer in K-fold cross-validation. Both models are strong candidates for predicting cost overruns in petrochemical projects in Iran.